# Fix MLX Metal Insufficient Memory & Command Buffer Execution Failed on Apple Silicon

> **Direct Error Triage**: On Apple Silicon Macs (M1/M2/M3/M4/M5), macOS enforces a default kernel cap that restricts Metal GPU allocations to roughly 75% of physical unified memory (e.g. ~18GB on a 24GB Mac, ~36GB on a 48GB Mac). When mlx_lm, Ollama, or llama.cpp crosses that wired-memory ceiling, Metal aborts with kIOGPUCommandBufferCallbackErrorOutOfMemory. Fix it instantly by raising iogpu.wired_limit_mb via sysctl to 85%–90% of physical RAM and enabling 4-bit/8-bit KV-cache quantization.

## Error Signature
```
RuntimeError: [metal::Device] Unable to allocate memory / Command buffer execution failed: Insufficient Memory (00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory)
```

## Root Cause & Memory Allocation Mechanics
Unlike discrete PCIe graphics cards with dedicated GDDR memory, Apple Silicon shares a single LPDDR5/LPDDR5X pool between macOS kernel tasks, WindowServer, and the Metal GPU. To prevent a runaway GPU compute shader from starving the XNU kernel and freezing the system, Apple's `IOAccelerator` driver rejects buffer allocations once pinned ('wired') GPU pages exceed `sysctl iogpu.wired_limit_mb` (which defaults to `0`, meaning automatic ~75% dynamic limit). Furthermore, long-prompt prefill passes in `mlx-lm` or `llama.cpp` without `--kv-bits 8` can spike transient Metal command-buffer memory past the wired threshold, triggering error `00000008:kIOGPUCommandBufferCallbackErrorOutOfMemory`.

## Step-by-Step CLI Remediation
### Ollama
```bash
# Raise Metal wired memory limit on a 48GB Mac to 42GB (43008 MB) or 64GB Mac to 56GB (57344 MB)
sudo sysctl iogpu.wired_limit_mb=43008
export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0
ollama serve
```

### Llama.cpp
```bash
# Cap prompt batch size (-b 512 -ub 256) to prevent Metal command buffer timeout/OOM
sudo sysctl iogpu.wired_limit_mb=43008
./llama-server -m model.gguf -ngl 99 -fa on -ctk q8_0 -ctv q8_0 -b 512 -ub 256
```

### vLLM
```bash
# Native Apple MLX-LM CLI with quantized 8-bit KV cache and capped KV size:
mlx_lm.generate --model mlx-community/Llama-3.3-70B-Instruct-4bit \
  --kv-bits 8 \
  --max-kv-size 16384 \
  --prompt "Explain unified memory limits"
```

## Verified Hardware & Cloud Pathways
- Explore 24GB GPUs on Amazon: https://www.amazon.com/s?k=RTX+4090+24GB+graphics+card&tag=sparecables-20 (rel="noopener sponsored")
- Rent on RunPod: https://runpod.io?ref=ynypzu94 (rel="noopener sponsored")
- Rent on Vast.ai: https://cloud.vast.ai/?ref_id=723130 (rel="noopener sponsored")

---
*Interactive Sizing Calculator available at: https://www.gpugarden.com/oom/mlx-metal-out-of-memory/*